Jiayu (Mila) Wang

EvoHarnessBench: Can Your Agents Keep Pace with an Evolving Harness?New

A benchmark for agents whose tools, skills, and specialist agents keep changing — showing that harness expansion alone can make agents forget tasks they used to solve.

SkillOrchestra: Learning to Route Agents via Skill Transfer

Learns transferable skills and per-model competence profiles to route compound AI systems — 700× cheaper to learn than Router-R1.

MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks

LiveResearchBench: A Live Benchmark for User-Centric Deep Research in the Wild

A live benchmark for deep research, with an evaluation suite that closely aligns with human judgment across six dimensions.

Synthesizing Agentic Data for Web Agents with Progressive Difficulty Enhancement Mechanisms

Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning

SPARKLE: what exactly does RL improve? Dissects reasoning into plan-following, problem decomposition, and knowledge use — and shows how to turn hard problems into training signal.

Helpful Agent Meets Deceptive Judge: Understanding Vulnerabilities in Agentic Workflows

COSMOS: Predictable and Cost-Effective Adaptation of LLMs

Accurately and efficiently predicts the performance and cost of training- and test-time adaptation strategies — smarter choices, fewer experiments.

Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models

Grammar-Aligned Decoding